DE102019214640A1 - CONTROL DEVICE AND CONTROL METHOD - Google Patents

Abstract

A control device that controls a controlled object includes a control execution device that applies a control output to the controlled object according to a control rule given to it, and a control method learning device that evaluates the control output applied to the controlled object using a specified evaluation function Using an evaluation result of the same learning data that learns learning data to thereby create the control rule and applies the control rule to the control execution device, and an evaluation function setting part that includes a plurality of evaluation functions in advance, one of the many evaluation functions based on selects a control state for the controlled object and communicates the selected evaluation function to the control method learning device.

Description

BACKGROUND OF THE INVENTION

FIELD OF THE INVENTION

The present invention relates to a method for performing real-time feedback control using artificial intelligence such as a neural network or the like.

DESCRIPTION OF THE PRIOR ART

Plant controls based on different control theories have previously been carried out in different plants in order to obtain a desired control result by controlling them.

In control, for example, of a rolling mill serving as an example of a plant, fuzzy control or neuro-fuzzy control has been applied as a control theory aimed at shape control to control, as an example of control, a corrugation state of a sheet. The fuzzy control was applied to a shape control using a coolant, and the neuro-fuzzy control was applied to a shape control of a Sendzimir mill. Of these, the shape control to which the neuro-fuzzy control was applied was performed by, as shown in Patent Document 1, determining a degree of similarity between the difference between an actual shape pattern detected by a shape detector and a target shape pattern and a preset one Reference shape pattern and determination of a control output variable with respect to an end of operation are carried out from the degree of similarity by a control rule represented by a control operation end operation variable with respect to the reference shape pattern likewise set in advance. It is assumed that the shape control of the Sendzimir rolling mill using the neuro-fuzzy control is used below as the state of the art.

1 illustrates the shape control of the in 1 Sendzimir rolling mill described by Patent Document 1. The neuro-fuzzy control was used in the shape control of the Sendzimir rolling mill. In this example, a pattern recognition mechanism performs 51 a shape pattern recognition on a by a shape detector 52 and then calculates whether the actual shape comes closest to one of the reference shape patterns set in advance. The control calculation mechanism 53 performs control using one of control operation end operation quantities with respect to such preset shape patterns as in FIG 2nd configured control rule shown. How regarding 2nd more specifically described, the pattern recognition mechanism computes 51 whether a difference ( Δε ) between that by the shape detector 52 recorded actual shape result and a target shape ( εref ) one of shape patterns ( ε ) 1 to 8 comes closest. The control calculation mechanism 53 selects and leads one of the tax procedures 1 to 8th out.

However, in the method in Patent Document 1, there is shown a case in which a change in shape contrary to expectations occurs, although there is shown a case in which an operator is allowed to manually operate to check the control rule while rolling, whereby the control rule or the like is checked. That is, a case occurs in which the control rule determined in the above-described manner is actually not compliant. This is due to a lack of studies of mechanical properties, an operating condition of a rolling mill and a change in the mechanical condition. It is difficult to check one by one whether the control rule set in advance is the best rule because many conditions have to be taken into account. Therefore, once the control rule is set, it is often used unchanged if it is not defective.

Since the control rule is set where the control rule is actually non-compliant due to a change in the operating condition or the like, it becomes difficult to improve the control accuracy to some extent or further. Furthermore, since the operator does not perform manual operation (malfunction occurs for control) as soon as the mold control is operated, it is also difficult to find a new control rule by manual intervention by the operator. Furthermore, even if a rolling material is rolled with a new standard, it is difficult to set the control rule according to the respective material.

In the conventional shape control as described above, since the control is performed using the control rule set in advance, there is a problem in that it is difficult to correct the control rule.

1. 1.) während der Ausführung der Formsteuerung während des Walzens eine neue Steuerungsregel finden und
2. 2.) da eine neue Steuerungsregel nicht vorab vorausberechnet werden kann und eine vollkommen unberechenbare Steuerungsregel optimal gemacht werden kann, wird das Steueroperationsende aufs Geratewohl ausgeführt, um bei Betrachtung desselben ein darauf bezogenes Steuerungsergebnis zu ermitteln.

In order to solve this problem, the following items 1.) and 2.) as shown in Patent Document 2 were implemented by changing the control rule at random while performing shape control and learning a rule in which the shape becomes good:

1. 1.) find a new control rule during the execution of the shape control during rolling and
2. 2.) since a new control rule cannot be calculated in advance and a completely unpredictable control rule can be made optimal, the end of the control operation is carried out at random in order to determine a control result related to it.

LIST OF PRINTINGS

• Patent document 1: Japanese Patent No. 2804161
• Patent document 2: Japanese Patent No. 4003733

BRIEF DESCRIPTION OF THE INVENTION

In the above prior art, a typical shape is set in advance as the reference shape pattern, and control is performed based on the control rule indicative of the relationship of the control operation end operation quantity to the reference waveform pattern. The learning of the control rule also relates to the control operation end operation amount with respect to the reference waveform pattern. The typical pre-defined reference shape pattern is used unchanged. Therefore, there arises a problem in that the controller becomes a shape controller that only reacts with a certain shape pattern.

The reference shape pattern is defined by the human knowledge in advance of the rolling mill and accumulated experience of a shape result and a manual operation. However, it is difficult to cover all the shapes produced in a rolling mill and a rolled material to be controlled. Therefore, when a shape other than the reference shape pattern occurs, the control based on the shape control is not carried out and a shape deviation remains without being suppressed. Alternatively, there is also a case where the reference shape pattern is erroneously recognized as a similar reference shape pattern and an incorrect control operation is performed, so that the shape deteriorates conversely.

Therefore, the conventional shape control has a problem in that since the control rule is learned using the preset reference shape pattern and the associated control rule and the control is performed, there is a limit to improving the control accuracy.

To solve this, it is assumed that a combination pattern of actual result data of a control object plant recognizing plant control device is used to perform control on the control object plant, which is a control method learning device, which is a combination of the actual result data of the control object -Level and a control operation learns, and a control execution device, which performs control of the control object plant according to the combination of the learned actual result data and the control operation, characterized in that the control execution device a control rule execution part, which a Tax output according to the defined combination of the actual result data of the control object system and the control operation, a control output determination part which determines the admissibility / inadmissibility of the tax output output from the control rule execution part determines and the control method learning device reports that the actual result data and the control operation are incorrect, and a control output suppressing part, which, when the control output determination part has determined, that the actual result data of the control object system is given to the control output when the control output is issued Deteriorate control object attachment, prevents the control output from being output to the control object attachment, and the control method learning device includes a “control result good / faulty” determination part, which if the control execution device actually does the control output to the control object After a time delay until a control effect is shown in the actual result data, the system outputs a determination as to whether the control result is error-free or faulty as to whether the Actual result data improves or deteriorates compared to those before the corresponding control, carries out a learning data creation part which, using the error-free / incorrectness of the control result in the “control result good / faulty” determination part and the control output, teaches data, and a control rule Learning part, which learns the actual result data and the teaching data as learning data, and characterized in that with the learning by the control method learning device, a combination of individual actual result data and a control operation for a plurality of controlled objects according to the state of the control object System and the combination of actual result data and the control operation thus obtained is used as a defined combination of actual result data of the control object system and a control operation in the control rule execution part.

It is very important that an evaluation function used to determine the control result as "good / faulty" is suitable. When determining the evaluation function, however, a developer of the control device subjectively determines the evaluation function while he is interviewing an operating engineer, an operator, etc. of the control object system and confirming the operation of a specific system. There are many cases in which it is unknown whether the evaluation function is set correctly and appropriately.

Let us consider a form control of a rolling mill as an example. It is ideal if, when controlling the shape of the rolling mill, a target shape and an actual shape coincide with one another in one direction of a sheet width. However, there are many cases in which they actually do not match. It is therefore common for a certain area of a sheet to be given special attention in practical work and for the actual shape and the desired shape to be controlled so that they coincide in the area. As an evaluation function for evaluating the shape of the sheet, an evaluation function is used in which each part in the direction of the sheet width is weighted in accordance with a shape deviation (= shape actual result - target shape) in each part in the direction of the sheet width.

In the rolling mill, a control operation end with respect to the shape of an end (sheet end) in the direction of the sheet width is separated from a control operation end with respect to a part (middle part) other than this. However, there are many cases in which they affect each other. Furthermore, the shape deviation is often large because the sheet end is not held on both sides like the middle part. When a control is applied to the sheet metal end in the direction of the sheet width, its influence is exerted on the central part, so that the shape of the central part deteriorates or the reverse occurs. Thus, it is difficult to control the shapes of the sheet end and the middle part so that they coincide with a target value at the same time. In many cases, the operator gives priority to either the end of the sheet or the middle section and controls it manually.

If the evaluation function used in the “good / faulty” determination with respect to the control result carries out an evaluation that differs from the operator's imagination, the operator terminates the shape control operation by the control device and carries out a manual operation according to his own imagination. In this case, the shape control by the control device and the manual operation performed by the operator are in a competing state. As a result, it is also assumed that the operator switches off the mold control by the control device, which interferes with the manual operation by the operator. It is to be feared that in the event of a repetition, the operator does not switch on the mold control by the control device from the beginning.

If the evaluation function applied to the “good / faulty” determination with regard to the control result is set in such a way that it carries out an evaluation that is suitable for the operator's idea, it is possible not only for the control device to compete with the manual operation of the operator to reduce, but also to reduce the execution of manual operation by the operator. A burden imposed on the operator is reduced and an improvement in the accuracy of the shape control is also expected.

An object of the present invention is to provide a method which enables control to be carried out on the basis of a suitable “good / faulty” determination with respect to a control result.

A control device of the present disclosure is a control device that controls a controlled object. The control device includes a control execution device which applies a control output to the controlled object according to a control rule given to it, a control method learning device which controls the control output applied to the controlled object Evaluates use of a specified evaluation function, created by using an evaluation result of the same learning data, and the learning data to thereby create the control rule and applies the control rule to the control execution device, and an evaluation function setting part that includes a plurality of evaluation functions in advance which selects many evaluation functions based on a control state for the controlled object and notifies the selected evaluation function to the control method learning device.

According to the present disclosure, it is expected that control can be performed based on a suitable “good / faulty” determination regarding a control result.

Figure list

• 1 is a drawing showing a shape control of an in 1 Sendzimir rolling mill described by Patent Document 1 shows.
• 2nd Fig. 12 is a drawing showing a control rule configured for each shape pattern by a control operation end operation quantity.
• 3rd FIG. 12 is a drawing showing an outline of a plant control device according to an embodiment.
• 4th Fig. 12 is a drawing showing a specific example of a control rule execution part 10th according to the embodiment.
• 5 Fig. 12 is a drawing showing a specific example of a control rule learning part 11 according to the embodiment.
• 6 Fig. 12 is a block diagram showing an internal configuration of an evaluation function setting part 17th shows.
• 7 Fig. 12 is a drawing showing a configuration of a neural network when used to control the shape of the Sendzimir mill.
• 8th Fig. 12 is a drawing for explaining a shape deviation and a control method.
• 9 Fig. 12 is a drawing showing an overview of a control input data creation part 2nd shows.
• 10th Fig. 12 is a drawing showing an overview of a tax expenditure calculation part 3rd shows.
• 11 Fig. 12 is a drawing showing an example of transition of a rolling speed in the rolling mill.
• 12th Fig. 4 is a drawing showing an example of an evaluation function database DB5 shows.
• 13 Fig. 12 is a drawing for explaining an operation overview of an evaluation function selection process learning unit 173 .
• 14 Fig. 12 is a drawing for explaining an operation overview of an evaluation function learning unit 174 .
• 15 Fig. 12 is a drawing for explaining an outline configuration of the evaluation function learning unit 174 .
• 16 Fig. 12 is a drawing for explaining an overview of a tax expenditure determination part 5 .
• 17th Fig. 12 is a drawing for explaining an operation overview of a "control result good / faulty" determination part 6 .
• 18th Fig. 12 is a drawing for explaining an operation outline of a learning data generation part 7 .
• 19th Fig. 14 is a drawing showing a processing stage and a processing content in the learning data generation part 7 shows.
• 20th is a drawing which is in a learning data database DB2 stored data example shows.
• 21 Fig. 12 is a drawing showing an example of a neural network management table TB shows.
• 22 is a drawing which is an example of the learning data database DB2 shows.

DESCRIPTION OF THE EMBODIMENT

First, the discoveries in the present invention and the processes leading to the present invention will be described, taking shape control of a rolling mill as an example.

First, the following three points are required to solve the above problems.

1.) A reference shape pattern and a control operation corresponding thereto are individually set in advance, the combination of the shape pattern and the control operation is learned without learning a control operation method, and the control operation is performed using the same.

2.) Since a new control rule cannot be calculated in advance and a completely unpredictable control rule can be made optimal, the control operation end is carried out at random in order to determine a control result related to it.

3.) An evaluation function is selected in accordance with the condition of the rolling mill with regard to freedom from errors / faultiness of the control result, and the selection of a suitable control rule is made possible.

In order to realize these three points, the control operation can preferably be changed so that the control result improves while changing the combination of the shape pattern and the control operation used for the shape control. For this purpose, the combination of the shape pattern and the control operation suitable for the shape pattern is learned by artificial intelligence such as a neural network or the like, and the output of the control operation can be preferably changed with respect to the shape pattern generated in the rolling mill.

If the control operation is changed while performing shape control on the rolling mill during operation, an incorrect control output is output, so that the shape of the sheet deteriorates. An operational irregularity such as a sheet metal breakage can thus occur. When the sheet breakage occurs, it takes time to replace the rolls used in the rolling mill and a material to be rolled during the rolling must be discarded, causing great damage. Therefore, it is necessary to prevent, as far as possible, the wrong tax output from being output to the rolling mill. Therefore, the evaluation function for determining the correctness of the shape can be changed according to the rolling condition.

The rolling state denotes a rolling state in which the rolling mill to be controlled is located. If the controlled object is not limited to the rolling mill, the rolling condition is generalized and can be related to a control condition. The rolling state can be distinguished by various parameters such as the control operation applied to the rolling mill, the state of the rolling mill, the state of rolling by the rolling mill, etc. In the present embodiment, the rolling condition can be distinguished by a rolling speed, for example.

From the above, in the present embodiment, in order to realize the above, the correctness of the control operation output from the neural network is checked using, for example, a simplified model of the rolling mill. An output in which the shape is considered to be significantly deteriorated is not output to the control operation end of the rolling mill, thereby preventing the shape from deteriorating. With respect to the neural network, the learning is carried out on the assumption that the control operation is incorrect in its shape pattern.

Since there is a possibility that the method for checking the correctness / defectiveness of the control operation itself is wrong, a control operation output of a neural network which has been determined to be false with a certain probability is also output to the control operation end of the rolling mill, which also enables learn the combination of an unforeseen shape pattern and a control operation.

Embodiments of the present invention will now be described in detail below with reference to the drawings.

3rd shows an overview of a plant control device according to an embodiment. The plant control device in 3rd consists of a control object system 1 , a control Execution device 20th which actual result data Si from the control object system 1 receives and one according to a control rule (a neural network) as in 2nd shown defined control operation size output SO to the control object system 1 supplies to the control object facility 1 to control a control procedure learning device 21 which the actual result data Si or the like from the control object system 1 receives to perform the learning, and the learned control rule in the control rule in the control execution device 20th reflects a variety of databases DB ( DB1 to DB3 ) and an administration table TB for the databases DB .

The control execution device 20th contains a control input data creation part as main elements 2nd , a control rule execution part 10th , a tax expenditure calculation part 3rd , a tax expense suppressing part 4th , a tax expenditure determination part 5 and a control operation failure generation part 16 .

In the control execution device 20th becomes the control input data creation part 2nd first used to derive the actual result data Si of a rolling mill, which is the control object system 1 is input data S1 of the control rule execution part 10th to create. The control rule execution part 10th created using the neural network (the control rule), which (which) the relationship between the actual result data Si of the object to be controlled and a control operation end operation command S2 represents the control operation end operation command S2 from the actual result data Si of the object to be controlled. The tax expenditure calculation part 3rd calculated based on the control operation end operation command S2 a control operation size S3 for the tax operation end. Thus, the control operation size S3 using the neural network according to the actual result data Si of the control object system 1 created.

The tax expenditure also determines the determination part 5 in the control execution device 20th using the actual result data Si from the control object system 1 and the tax operation size S3 from the tax expenditure calculation part 3rd a control operation size output allowance / inadmissibility data value S4 for the tax operation end. The tax expense suppressing part 4th determines the admissibility / inadmissibility of the tax operation variable S3 for the control operation end according to the control operation quantity output allowance / inadmissibility data value S4 and gives the allowable tax operation size S3 than that to the control object system 1 Tax operation size output to be delivered SO out. Thus, the control operation size determined as irregular S3 not to the control object system 1 spent. Incidentally, the control operation disturbance generation part generates 16 a fault and sends it to the control object system for checking the system control device 1 .

The control execution device configured as described above 20th also accesses a control rule database DB1 and an output determination database DB3 back to perform its processing, as will be described later. The control rule database DB1 is both with the control rule execution part 10th in the control execution device 20th as well as with the control rule learning part 11 in the control procedure learning device 21 connected connected. A control rule (a neural network) as a learning result in the control rule learning part 11 is in the control rule database DB1 saved. The control rule execution part 10th accesses the in the control rule database DB1 saved control rule. The output determination database DB3 is with the tax expenditure determination part 5 in the control execution device 20th connected connected.

4th shows a specific example of the control rule execution part 10th according to the present embodiment. The control rule execution part 10th receives that in the control input data creation part 2nd created input data S1 and issues the control operation end operation command S2 to the tax expenditure calculation part 3rd . The control rule execution part 10th has a neural network 101 . In the neural network 101 is the control operation end operation command S2 basically through the procedure of the patent document 1 as in 2nd shown defined. In the present invention, the control rule execution part has 10th also via a neural network selection unit 102 which is the optimal control rule as the control rule in the neural network 101 selects, referring to those in the control rule database DB1 stored control rules and causes them to be executed. Thus, the control rule execution part chooses 10th in 4th he uses a required neural network from a plurality of neural networks, which are divided for the purposes of an operator group and control. The control rule database DB1 can preferably also such actual result data (data of an operation group or the like) Si that a neural network and a “good / faulty” determination reference can be selected as data from the control object system 1 contain. Incidentally, the neural network and the Control rule, since the neural network has a relationship that it serves as the control rule when it is executed, are not distinguished from each other and are used in the same sense.

As again in 3rd shown, the control method learning device 21 learning that in the control execution device 20th used neural network 101 out. If the control execution device 20th the control operation quantity output SO to the control object system 1 outputs, it takes time until a control effect actually shows up as a change in the actual result data Si. Therefore, the learning is carried out using data delayed by this time. in 3rd Z ^-1 specifies a suitable time delay function for all data.

The control procedure learning device 21 contains a "control result good / faulty" determination part as main elements 6 , a learning data creation part 7 , a control rule learning part 11 and an evaluation function setting part 17th .

Of these, the “result good / faulty” determination part is determined 6 using the actual result data Si and the previous actual result data Si0 from the control object system 1 and by the evaluation function setting part 17th set evaluation function, whether the actual result data Si change in an improvement direction or in a direction of deterioration, and it gives a "control result good / faulty" data value S6 out.

The learning data creation part 7 in the control procedure learning device 21 created using the by delaying the input data such as the control operation end operation command S2 , the control operation size S3 , the control operation size output allowance / inadmissibility data value S4 etc., each of which is simultaneously in the control execution device 20th were created, and the "control result good / faulty" data value S6 from the "control result good / faulty" determination part 6 Data obtained for learning the neural network used new teaching data S7a and delivers this to the control rule learning part 11 . Incidentally, the teaching dates correspond S7a from the control rule execution part 10th issued control operation end operation command S2 . One can say that the learning data creation part 7 by estimating that from the control rule execution part 10th issued control operation end operation command S2 using the determination part of the "control result good / faulty" 6 supplied "control result good / faulty" data value S6 received data as the new teaching data S7a has determined.

5 illustrates a specific example of the control rule learning part 11 according to the present embodiment. The control rule learning part 11 contains an input data creation unit as main components 114 , a teaching data creation unit 115 , a neural network processing unit 110 and a neural network selection unit 113 . Furthermore, the control rule learning part receives 11 by delaying the input data S1 from the input data creation part 2nd and new teaching dates S7a from the learning data creation part 7 data received S8a as external inputs and accesses those in the control rule database DB1 and the learning data database DB3 saved data.

In the control rule learning part 11 will be the input data S1 subjected to the appropriate time delay compensation and then via the input data creation unit 114 into the neural network processing unit 110 delivered.

Furthermore, in the control rule learning part 11 the new teaching dates S7a from the learning data creation part 7 as total teaching data S7c , which is also in the learning data database DB2 saved past teaching data S7b included, via the teaching data creation unit 115 to the neural network processing unit 110 delivered. This teaching data S7a and S7b are appropriately stored in the learning data database DB2 saved and used.

Likewise, the input data S8a from the control input data creation part 2nd as total input data S8c , which is also in the learning data database DB2 saved past input data S8b included, via the input data creation unit 114 to the neural network processing unit 110 delivered. This input data S8a and S8b are appropriately stored in the learning data database DB2 saved and used.

The neural network processing unit 110 consists of a neural network 111 and a neural network learning control part 112 . The neural network 111 captures the input data S8c from the input data creation unit 114 who have favourited Teaching Dates S7c from the teaching data creation unit 115 and the through the neural network selection unit 113 selected control rule (the neural network) and stores the finally determined neural network in the control rule database DB1 .

The neural network learning control part 112 controls the input data creation unit 114 , the teaching data creation unit 115 and the neural network selection unit 113 at appropriate times and controls them to input the neural network 111 to get and the processing result in the control rule database DB1 save.

Here are the neural network 101 in the control execution device 20th in 4th and the neural network 111 in the control procedure learning device 21 in 5 both neural networks with the same concept. A description now follows regarding the difference in the basic concept of where they are used.

The neural network 101 in the control execution device 20th is a neural network of a predefined content and is used to obtain the control operation end operation command S2 than the corresponding output when the input data S1 are supplied, a so-called neural network used for processing in one direction. On the other hand, the neural network serves 111 in the control procedure learning device 21 when the input data S8c and teaching dates S7c about the input data S1 and the control operation end operation command S2 are set as the learning data to determine a neural network that fulfills this input / output relationship by learning.

The basic idea of processing in the control method learning device configured as described above 21 is the following: First, when the content of the control operation size output allowance / inadmissibility data value S4 "Allowed" is the tax operation size output SO to the control object system 1 spent. If the content of the "control result good / faulty" data value S6 "Good" is (the actual result data Si change in the direction of improvement), it is determined that that from the control rule execution part 10th issued control operation end operation command S2 is correct, and the learning data is created so that the output of the neural network has the control operation end operation command S2 reached.

On the other hand, if the content of the control operation size output allowance / inadmissibility data value S4 "Illegal" is or the control operation size output SO to the control object system 1 is output and the content of the "control result good / faulty" data value S6 "Faulty" is (the actual result data Si change in the direction of deterioration), determined that that by the control rule execution part 10th issued control operation end operation command S2 is incorrect, and the learning data are created so that the output of the neural network is not delivered. Here, the neural network output is configured in such a way that two kinds of outputs in “+” and “-” directions are output as control outputs to the same control operation end, and the learning data are created so that the control operation end operation command S2 is not issued on its exit page.

Furthermore, in the 5 control rule learning part shown 11 the processing due to the data processing by the neural network learning control part 112 done in the following way. Here the learning in the control rule execution part 10th used neural network 101 first using learning data which is the combination of S8c , obtained by delaying the input data S1 for the control execution device 20th , and that in the teaching data creation unit 115 created teaching data S7c are executed. In fact, it's the same neural network 111 like the neural network 101 of the control rule execution part 10th in the control rule learning part 11 and is thereby designed to perform a functional test under various conditions in order to learn an answer at that time and then obtain a control rule under which it is confirmed that a better result is obtained as its learning result. Since it is necessary to enable the learning to be carried out using a large number of learning data, a large number of past learning data is stored in the learning data database DB2 , which has saved the learning data created in the past, and learned to carry out its processing. This is accompanied by the existing learning data in the learning data database DB2 saved. Furthermore, the learned neural network is in the control rule database DB1 stored to it in the control rule execution part 10th to use.

The neural network can be learned together using the past learning data each time the new learning data is created. After the learning data has been saved to a certain extent (for example 100 pieces), the neural network can be learned together using the past learning data.

Furthermore, the “control result good / faulty” determination part leads 6 using the by the evaluation function setting part 17th set evaluation function a "good / faulty" determination by. The “good / faulty” determination of the control result differs in the determination result according to the evaluation function to be used. For this reason, neural networks corresponding to a multiplicity of evaluation functions are created. As for the same input data, teaching data are created by the evaluation functions and then learned. A multitude of teaching data is thus created with respect to the one input data and used to learn the neural networks corresponding to the respective teaching data. At the same time, it is possible to learn the neural networks corresponding to the many evaluation functions. Here are the many evaluation functions for respective procedures, such as in the case of a shape control, for example which part (sheet end, middle part, asymmetrical part, etc.) should preferably be controlled in a direction of the sheet width or which of a large number of points to be controlled (e.g. sheet thickness and Tension, rolling pressure etc.) should preferably be controlled etc., evaluation functions used.

Applying the present embodiment, it is assumed that once that is in the control rule execution part 10th used neural network 101 learned that a new control operation is not performed. Therefore, a new operation method is arbitrarily appropriately adopted by the control operation failure generation part 16 generated to a control operation in addition to the control operation size S3 to carry out, whereby a new control method is learned.

In the following, the details of the present plant control method with the shape control in such a Sendzimir rolling mill as shown in Patent Document 1 as an object will be described as an example. Incidentally, in the description of the shape control, the following specifications A and B introduced.

The specification A is a specification of an evaluation function and presumably contains information about the priority in the direction of the sheet width. For example, in the shape control, there are many cases where it is difficult to control the sheet thickness or the like to a target value over the entire area in the direction of the sheet width with respect to mechanical properties. That is why there are evaluation functions in the direction of the sheet width A1 to ON ( N is the maximum number of evaluation functions to be set) according to the following many procedures.

The evaluation function is defined so that its value becomes small when the evaluation is good. For example, there is the root mean square of a control deviation, the maximum value - minimum value, etc.

• A1: Eine Auswertungsfunktion wird verwendet, bei welcher ein Blechende Priorität hat und die Gewichtung des Blechendes gewichtet wird. $$J_{A1}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{2}Randteilwc\left(i\right)=\mathrm{3,0,}Mittelteilwc\left(i\right)=\mathrm{1,0}}$$
  
• A2: Eine Auswertungsfunktion wird verwendet, bei welcher der Mittelteil Priorität hat und die Gewichtung des Mittelteils gewichtet wird. $$J_{A2}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{2}Randteilwc\left(i\right)=\mathrm{1,0,}Mittelteilwc\left(i\right)=\mathrm{3,0}}$$
  
• A3: Eine Ausdehnungsrichtung des Blechendes ist zulässig. $$\begin{array}{l}{J}_{A3}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^2\text{ w}c\left(i\right)=\mathrm{1,0}}\\ Randteil\epsilon \left(i\right)=\epsilon \left(i\right),\text{ w}enn\left(\epsilon \left(i\right)<0\right)\mathrm{,0,}wenn\left(\epsilon \left(i\right)\ge 0\right)\end{array}$$
  
• A4: Eine Expansionsrichtung des Blechendes ist zulässig. $$\begin{array}{l}{J}_{A4}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^2,wc\left(i\right)=\mathrm{1,0}}\\ Randteil\epsilon \left(i\right)=\epsilon \left(i\right),wenn\left(\epsilon \left(i\right)\ge 0\right)\mathrm{,0,}wenn\left(\epsilon \left(i\right)<0\right)\end{array}$$
  
• A5: Zulässig, wo das Blechende sich in einem Unempfindlichkeitsbereich befindet. $$\begin{array}{l}{J}_{A5}\left(\epsilon \left(i\right)\right)=\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^2,wc\left(i\right)=\mathrm{1,0}}\\ Randteil\epsilon \left(i\right)=\mathrm{0,}wenn\left(UL>\epsilon \left(i\right)>LL\right),sonst\epsilon \left(i\right)=\epsilon \left(i\right)\end{array}$$
  
• A6: Maximalwert - Minimalwert $$J_{A6}\left(\epsilon \left(i\right)\right)=max\left(\epsilon \left(i\right)\right)-min\left(\epsilon \left(i\right)\right)$$
  

Here it is assumed that there are six types of procedures and evaluation functions to be illustrated below A1 to A6 can be used as an example.

• A1: An evaluation function is used in which a sheet end has priority and the weighting of the sheet end is weighted. $$J_{A1}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{\mathrm{2nd}}Randteilwc\left(i\right)=\mathrm{3.0,}Mittelteilwc\left(i\right)=1.0}$$
  
• A2: An evaluation function is used in which the middle part has priority and the weighting of the middle part is weighted. $$J_{A\mathrm{2nd}}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{\mathrm{2nd}}Randteilwc\left(i\right)=\mathrm{1.0,}Mittelteilwc\left(i\right)=3.0}$$
  
• A3: An expansion direction of the sheet end is permissible. $$\begin{array}{l}{J}_{A\mathrm{3rd}}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{\mathrm{2nd}}\text{w}c\left(i\right)=1.0}\\ Randteil\epsilon \left(i\right)=\epsilon \left(i\right),\text{w}enn\left(\epsilon \left(i\right)<0\right)0wenn\left(\epsilon \left(i\right)\ge 0\right)\end{array}$$
  
• A4: An expansion direction of the sheet end is permissible. $$\begin{array}{l}{J}_{A\mathrm{4th}}\left(\epsilon \left(i\right)\right)=\frac{1}{\text{n}}{\displaystyle \sum _{\text{i}=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{\mathrm{2nd}},wc\left(i\right)=1.0}\\ Randteil\epsilon \left(i\right)=\epsilon \left(i\right),wenn\left(\epsilon \left(i\right)\ge 0\right)0wenn\left(\epsilon \left(i\right)<0\right)\end{array}$$
  
• A5: Allowed where the sheet metal end is in an insensitive area. $$\begin{array}{l}{J}_{A5}\left(\epsilon \left(i\right)\right)=\frac{1}{n}{\displaystyle \sum _{i=1}^n{\left(wc\left(i\right)\cdot \epsilon \left(i\right)\right)}^{\mathrm{2nd}},wc\left(i\right)=1.0}\\ Randteil\epsilon \left(i\right)=\mathrm{0,}wenn\left(UL>\epsilon \left(i\right)>LL\right),sOnst\epsilon \left(i\right)=\epsilon \left(i\right)\end{array}$$
  
• A6: maximum value - minimum value $$J_{A6}\left(\epsilon \left(i\right)\right)=max\left(\epsilon \left(i\right)\right)-min\left(\epsilon \left(i\right)\right)$$
  

6 Fig. 12 is a block diagram showing an internal configuration of the evaluation function setting part 17th shows. The evaluation function setting part 17th contains an evaluation function manual setting unit 171 , an evaluation function selection unit 172 , an evaluation function selection process learning unit 173 and an evaluation function learning unit 174 . The evaluation function setting part 17th performs the following processing of the evaluation function together with the evaluation function database DB5 out.

processing 17-1 : Setting the evaluation function

The evaluation function manual setting unit 171 sets an evaluation function. This is processing which mathematizes and sets ways of conceiving forms of an operational engineer and an operator in advance.

processing 17-2 : Selection of an evaluation function

The evaluation function selection unit 172 selects one in the control execution device 20th used evaluation function according to a rolling condition.

processing 17-3 : Learning the selection process of the evaluation function

The evaluation function selection process lesson 173 performs the learning in such a way that an evaluation function corresponding to the rolling condition is selected according to the rolling condition and a manual operation result of the operator.

processing 17-4 : Learning the evaluation function itself

Since the evaluation set in advance by manual operation is not necessarily correct, the evaluation function learning unit learns 174 the evaluation function itself. Here, the evaluation function to be learned is referred to as a learning evaluation function. As the learning progresses to a certain extent, the evaluation using the learning evaluation function is enabled. In this case, the learning evaluation function can be used for evaluation as the evaluation function.

The specification B is a specification about a correspondence to a previously known condition. As an example, since the relationship between the shape pattern and the control method changes depending on various conditions, it is assumed that there is a need, the specification B1 and the specification B2 for example, to subdivide the sheet width and the steel grade according to their categories. With the above respective changes, the degree of influence on the shape of the mold operation end changes.

In this example is the control object system 1 a Sendzimir mill and the actual result data becomes a form-actual result. Incidentally, the Sendzimir rolling mill is a rolling mill with a multiple roll for cold rolling a hard material such as stainless steel or the like. In the Sendzimir rolling mill, a small work roll is used for the purpose of rolling the hard material with high elongation. Therefore, it is difficult to get flat stainless steel sheet. As a countermeasure, the construction of the multiple roller and various shape control parts were introduced. The Sendzimir mill usually contains upper ones and lower first intermediate rolls, which are each conical and displaceable on one side, and is additionally vertical with six dividing rolls and two as AS-U designated rollers equipped. In examples to be described later, detection data of a shape detector is used as the actual result data Si of the shape, and furthermore, a shape deviation, which is a difference from a target shape, is used as the input data S1 used. It is also assumed that the control operation size S3 AS-U No. 1 to No. n and the roll displacement amounts of the upper and lower first intermediate rolls.

7 shows a configuration of a neural network when used to control the shape of the Sendzimir mill. The neural network can be briefly referred to as a neural network. Here the neural network designates the neural network 101 for the control rule execution part 10th and designates it through the neural network 111 for the control rule learning part 11 shown neural network, but both are the same in structure.

In the example of the shape control of the Sendzimir rolling mill shown in the present embodiment, the actual result data Si are from the control object system 1 the data (here, it is assumed that the shape deviation, which is the difference between the actual result shape and the target shape, is output) of the shape detector containing actual result data of the Sendzimir mill. The control input data creation part 2nd receives a standardized form deviation 201 and a shape deviation level 202 than the input data S1 . Thus each contains input layers of the neural networks 101 and 111 the standardized form deviation 201 and the shape deviation level 202 . Incidentally, the neural network, although in 7 the shape deviation level 202 when the input to the neural network input layer is set, be switched according to the stage.

An output layer also contains an AS-U actuation level 301 and a first intermediate degree of actuation 302 according to AS-U and the first intermediate roll, each of which is the form control operation end of the Sendzimir mill. In terms of AS-U their degree of operation has an AS-U opening direction (direction in which a roll gap (distance between vertical work rolls of the rolling mill) is opened) and an AS-U closing direction (direction in which the roll gap is closed) for each AS-U . Further, with regard to the first intermediate roller, its degree of operation has a first intermediate roller opening direction (direction in which the first intermediate roller is operated outward from the center side of the rolling mill) and a first intermediate roller closing direction (direction in which the first intermediate roller is operated toward the center of the mill) for each vertical first intermediate roller. For example when the shape detector 20th Contains areas and the shape deviation level 202 is set to three levels (large, medium and small), the input layer changes to 23 Entrances. Furthermore, the starting layer, if the AS-U has seven saddles and it is assumed that the vertical first intermediate rolls are capable of shifting in the direction of the sheet width, fourteen AS-U actuation degrees 301 and four first intermediate degrees of operation, which is a total 18th results. The number of layers of intermediate layers and the number of neurons of each layer are set appropriately. Incidentally, although later based on 10th described, with respect to the form control operation end of the Sendzimir mill, which is the output layer, the neural network output is configured to output two kinds of outputs in the "+" and "-" directions at each individual control operation end .

8th shows the shape deviation and the control method. Here is in the upper part of 8th shown a control method in which the shape deviation is large. In the lower part of 8th a control method is shown in which the shape deviation is small. Incidentally, the vertical direction indicates the magnitude of the shape deviation, the horizontal axis direction indicates the direction of the sheet width, both sides of the sheet width indicate a sheet end and the middle indicates a sheet metal middle part. As in the upper part of 8th shown, with a large shape deviation, the entire shape is changed more preferentially than the local shape deviation in the direction of the sheet width. On the other hand, as in the lower part of 8th shown, the local shape deviation is preferably reduced in the case of a small shape deviation.

Thus, since it is necessary to change the control method according to the magnitude of the shape deviation, as in FIG 7 shown the shape deviation level 202 provided and on the neural networks 101 and 111 applied to determine the magnitude of the shape deviation. With regard to the shape deviation, for example, those standardized to 0 to 1 can be used regardless of the magnitude of the shape deviation. This is an example, and it is also contemplated to enter it unchanged into the input layer of the neural network without standardizing the shape deviation. In addition, it is assumed that the neural network itself is changed according to the magnitude of the shape deviation (for example, two neural networks are provided and one in one large shape deviation used neural network and a neural network used for a small shape deviation)).

The neural networks 101 and 111 , each of which has such a configuration as that described above 7 shown is allowed to learn the operation method for the shape pattern, and shape control is carried out using the neural networks under learning. The neural networks having the same configuration also differ in their properties depending on a learning condition and are capable of outputting different tax expenditures with respect to the same shape pattern.

Therefore, the optimal control can be configured by properly using the many neural networks depending on the other conditions of the shape result with respect to different conditions. This corresponds to the specification B . The configuration of the 4th gives a specific example in which such a specification is made. In the example of the configuration in 4th a neural network other than that in the control control execution part becomes, according to rolling actual results, a rolling mill operator name, a steel grade of a material to be rolled, a sheet width thereof, etc. 10th used neural network 101 provided and in the control rule database DB1 registered. The neural network selection unit 102 selects a neural network that matches the condition at that time and provides this for the neural network 101 of the control rule execution part 10th on. Incidentally, as the condition at that time in the neural network selection unit 102 , Data about the sheet width from the actual result data Si in the control object system 1 removed, and a neural network is preferably selected according to the data. Furthermore, if the many neural networks used herein, such input and output layers as in 7 have shown, the number of layers of intermediate layers and the number of units of each layer differ from one another.

9 shows an overview of the control input data creation part 2nd which is the one in the input layer of each of the neural networks 101 and 111 data to be entered S1 (standardized form deviation 201 and shape deviation level 202 ) created. Here is the shape detector data of the shape detector, which is a sheet metal shape during rolling in which the control object system 1 representing the Sendzimir rolling mill first recorded a shape deviation PP value (peak-to-peak value) as the actual result data Si used as the input. Spp , which is the difference between the maximum and minimum values of the detection result of each shape detector area by a shape deviation PP value calculation unit 210 determined. A shape deviation level calculation unit 211 divides the shape deviation according to the shape deviation PP value Spp in three levels "large", "medium" and "small". The shape is a distribution of an expansion ratio of the rolled material in the direction of the sheet width, and I-UNIT , which gives the elongation ratio in 10-5 units, is used as the unit. For example, the shape deviation is divided according to the following equations.

$$50I-UNIT>S_{PP}\ge 10I-UNIT$$

$$10I-UNIT>S_{PP}$$

Here, the shape deviation is divided in such a way that the shape deviation level becomes by the equation (1) (large = 1, medium = 0 and small = 0), the shape deviation level by the equation (2) (large 0, medium = 1 and small = 0) and the shape deviation level becomes by setting up equation (3) (large = 0, medium = 0 and small = 1). Incidentally, the shape deviation of each area is standardized here using S _PM with S _PM = S _PP . $$S_{PP}\ge 50\mathrm{I.}-UN\mathrm{I.}T$$

$$50\mathrm{I.}-UN\mathrm{I.}T>S_{PP}\ge \mathrm{10th}\mathrm{I.}-UN\mathrm{I.}T$$

$$\mathrm{10th}\mathrm{I.}-UN\mathrm{I.}T>S_{PP}$$

The standardized form deviation 201 and the shape deviation level 202 which the input data into the neural network 101 are created in the manner described above. The standardized form deviation 201 and the shape deviation level 202 are the input data S1 of the control rule execution part 10th .

10th shows an overview of the tax expenditure calculation part 3rd . The tax expenditure calculation part 3rd creates a control operation size S3 which is an operation command for each shape control operation end according to the control operation end operation command S2 (according to the AS-U Degree of activity 301 and the first intermediate degree of operation 302 in the example of the shape control in the Sendzimir mill) in the control control execution part 10th which is the output from the neural network 101 is. By the way, here is every data example for the degree of AS-U actuation 301 and the first intermediate degree of operation 302 , which exist in many forms, shown and all data include a pair of opening direction degree and closing direction degree data.

Since the entered AS-U degree of actuation 301 The expenditure in the respective AS-U opening and closing directions multiplied by the tax expenditure calculation part 3rd the difference between them with a conversion gain GASU in it to thereby issue an operation command to each AS-U to spend. Because the tax expense to everyone AS-U becomes an AS-U position change amount (the unit of which is length), the conversion gain becomes a conversion gain from the degree to the position change amount.

Furthermore, since the first intermediate operation degree entered in the same manner 302 Has expenditure on the first intermediate outside and inside, the difference between the expenditure with a conversion gain G _1ST multiplied to thereby issue an operation command to each first intermediate roll shift. Since the control output to every first intermediate roller becomes a first intermediate roller shift position change amount (the unit of which is length), the conversion gain becomes G _1ST a conversion gain from the degree to the position change amount.

Thus, the control operation size S3 be calculated. The tax operation size S3 includes AS-U position change quantities No. 1 to No. n (n corresponds to the saddle number of the AS-U roller), an upper first intermediate displacement position change quantity and a lower first intermediate displacement position change quantity. Illustrated by the way 10th a system in which disturbance data from the control operation disturbance generation part 16 to the control operation end operation command S2 be added.

Now using 6 an operation overview of the evaluation function setting part 17th described. The evaluation function is one in which the operator's intention to control the shape in the rolling mill is reflected. The operator's intention changes according to the rolling condition. Here it is assumed that the rolling condition is distinguished by the rolling speed. As in 11 As shown, the rolling speed of the rolling mill changes in such a manner that the rolling speed is increased from its standstill state to carry out the rolling at a constant speed and the rolling speed is decreased to stop the rolling. The rolling condition also changes how 17-1 , 17-2 , 17-3 ... according to a change in the rolling speed. Then the operator's intention changes like an intention 1 , an intention 2nd , an intention 3rd , ... according to a change in the rolling condition. For example, the operator's intention may include those shown below.

• Intent 1: At the beginning of rolling at a low speed, a central part of a sheet is given priority to ensure the stability of its passage.
• Intent 2: When the rolling is accelerated, special attention is paid to a sheet end to prevent the sheet or the like from meandering.
• Intent 3: At a constant rolling speed, the quality of a material to be rolled is taken into account and a shape deviation of the sheet end in its direction of expansion is permitted and the shape of the middle part is given priority so that no sheet breakage is generated.

• Die Auswertungsfunktion A2 entspricht der Absicht 1.
• Die Auswertungsfunktion A1 entspricht der Absicht 2.
• Die Auswertungsfunktion A3 entspricht der Absicht 3.

If one links the above respective intentions with evaluation functions A1 to ON , the following results:

• The evaluation function A2 corresponds to the intention 1 .
• The evaluation function A1 corresponds to the intention 2nd .
• The evaluation function A3 corresponds to the intention 3rd .

The corresponding relationship between the operator's intentions described above and the evaluation functions is recorded in the evaluation function database DB5 saved. An example of the evaluation function database DB5 is in 12th shown. It is determined which of the evaluation functions A1 to A6 (Evaluation function NO ) should be used for any intention of the operator corresponding to the above rolling condition.

Because the rolling condition on which the intentions 1 , 2nd and 3rd applied, by which the rolling speed can be distinguished, it is possible to select whether one of the evaluation functions A1 to ON to be used according to the rolling speed. The operator or the operational engineer or the like makes the association between the rolling speed and the evaluation speeds A1 to ON in the evaluation function database DB5 using the evaluation function manual setting unit 171 by hand. The evaluation function selection unit 172 selects the evaluation function corresponding to the intention of the operator corresponding to the rolling condition defined by the actual rolling result Si (which contains the actual result of the rolling speed) and sets it for the control output determination part 5 and the "control result good / faulty" determination part 6 on.

The manual setting for the selection of the evaluation function by the operator or the operational engineer can differ in practice because of the cases in which the determination of a specific operator is not set correctly and the operator finds and uses a new determination criterion. In order to evaluate a “good / faulty” determination with regard to the manual setting, the evaluation function selection method learning unit determines 173 on the basis of the actual result data obtained by actual rolling operation, whether a method for selecting the evaluation function is good or defective. Furthermore, the evaluation function selection method learning unit changes 173 if it is determined that the selection method is faulty, the setting of the selection method of the evaluation function in the evaluation function database DB5 .

13 Fig. 12 is a drawing for explaining an operation overview of the evaluation function selection method learning unit 173 . If it is determined during the rolling operation that the shape of the sheet is defective, the operator starts manual operation and continues the manual operation until it is determined that the shape is good. Thus, the operator's intention is reflected when the operator starts the manual operation and ends the manual operation. The evaluation function selection process lesson 173 calculates form evaluation results for the respective evaluation functions A1 to ON based on data at that time. Comparing these shape evaluation results with each other is the evaluation function selection process learning unit 173 able to determine the relative quality of the evaluation function, ie which evaluation function comes close to the intention of the operator.

Assuming that the shape is shown to be satisfactory when a shape evaluation value becomes small, it can be determined that the evaluation function in which the shape evaluation value is large at the time the manual operation starts and the shape evaluation value at the time of the completion of the Manual operation is small, is an evaluation function suitable in its rolling state (at its rolling speed).

worin a ein Formauswertungswert bei Beginn der Handbetätigung und b ein Formauswertungswert bei Beendigung der Handbetätigung ist. Die Auswertungsfunktions-Auswahlverfahrens-Lerneinheit 173 ermittelt die Auswertungsfunktion Ai der Auswertungsfunktionen A1 bis AN, bei welcher das Verhältnis Xi, welches der „Auswertungsfunktion gut/fehlerhaft“-Ermittlungsindex ist, den größten Wert annimmt, als eine Auswertungsfunktion, bei welcher man eine mit der Absicht des Bedieners bei dem Walzzustand zu dieser Zeit am besten übereinstimmende Auswertung erhält, und wählt dieselbe als die beste Auswertungsfunktion aus.In the present embodiment, since the calculation method differs for each evaluation function such as the evaluation using the root mean square, the evaluation function using the maximum value or the minimum value, etc., it is required that a common index be used as an index ("evaluation function good / defective “-Determination index) is used and then compared with it to evaluate whether the evaluation function is good or faulty. Here, the evaluation function selection process learning unit compares 173 as an example, the evaluation functions using a ratio expressed by the following equation Xi : $$\text{Ratio Xi}=\left(\text{a}-\text{b}\right)/\text{b}$$

where a is a shape evaluation value at the start of manual operation and b is a shape evaluation value at the end of manual operation. The evaluation function selection process lesson 173 determines the evaluation function Ai of the evaluation functions A1 to ON at which the ratio Xi , which is the "evaluation function good / faulty" determination index, takes the largest value as an evaluation function in which an evaluation which best matches the intention of the operator in the rolling condition at that time is obtained, and selects it as the best evaluation function out.

The rolling condition at the beginning or end of manual operation and the intention of the operator at that time can be determined from rolling actual results. If the evaluation function assigned to the corresponding intention of the operator in the evaluation function database DB5 is a different evaluation function than the one selected here, updates the evaluation function selection process learning unit 173 the evaluation function according to the corresponding intention of the operator on the Evaluation function Ai. Then, the evaluation function selection process learning unit 173 the evaluation function Ai according to the settings after the change from the next time for the tax expenditure determination part 5 and the "control result good / faulty" determination part 6 on.

A graph in 13 shows the temporal transition of the shape evaluation values of the two evaluation functions A1 and A2 . The shape evaluation value of the evaluation function A1 at the time of manual actuation A1S , and the shape evaluation value at the time of the end of the manual operation A1E . The shape evaluation value of the evaluation function A2 at the time of manual actuation A2S , and the shape evaluation value at the time of the end of manual operation is A2E.

How from 13 is a ratio X2 = (A2S-A2E) / A2E of the evaluation function A2 greater than a ratio X1 = (A1S-A1E) / A1E of the evaluation function A1 .

Furthermore, the evaluation function learning unit leads 174 the learning of the evaluation function, taking into account also the possibility that the manually set evaluation function itself is not suitable.

14 Fig. 12 is a drawing for explaining an operation overview of the evaluation function learning unit 174 . The evaluation function lesson 174 receives as inputs an actual shape result, which is an actual result value of the shape of the sheet obtained by rolling, and an actual rolling result, which is a parameter value for a control operation during the rolling, installs a neural network (neural network for evaluation function) for the Evaluation function, which outputs a shape evaluation value, and carries out the learning of the neural network for the evaluation function using actual result data. Incidentally, such a rolling actual result (eg rolling speed) can be selected as the rolling actual result used as the input into the neural network for the evaluation function with regard to the influence of the evaluation function. The learned neural network can be used as the evaluation function.

As described above, regarding the evaluation of the shape that the operator intends, it can be said that the shape of the sheet is defective when the operator starts manual operation and the shape of the sheet is good when the operator ends manual operation. Thus, the evaluation function learning unit 174 in the process of producing the sheet by the rolling mill, the shape evaluation value at the time of the start of the manual operation by the operator is 1 (1 means that the shape is defective) and the shape evaluation value at the time of the end of the manual operation is 0 (0, that the shape is good) and stores it as teaching data together with a shape actual result and a rolling actual result at that time. Then the evaluation function learning unit leads 174 monitored learning of the neural network using the stored teaching data. Thus, since the shape evaluation value is output when the rolling actual result and the shape actual result are input, the learned neural network can be used as the evaluation function.

An overview configuration of the evaluation function learning unit 174 is in 15 shown. The tax expenditure determination part 5 and the "control result good / faulty" determination part 6 initially use the evaluation function set manually by the operator. The evaluation function lesson 174 Adds teaching data to be described later to the actual rolling result data Si containing the form actual result and the rolling actual result and learns their result as learning data in order to thereby create a neural network for the evaluation function, which provides an evaluation function instead of the initial evaluation function.

The evaluation function lesson 174 contains an evaluation execution part and a learning execution part.

The evaluation execution part contains a neural network 1740 for the evaluation function which is in the tax expenditure determination part 5 and the "control result good / faulty" determination part 6 is used and performs the evaluation using the neural network 1740 for the evaluation function.

The learning execution part contains a neural network 1741 for an evaluation function which corresponds to the neural network 1740 is equivalent to the evaluation function, and leads learning using the neural network 1741 for the evaluation function. Here's how in 14 shown the neural network 1741 a neural network for the evaluation function, which contains an actual shape result and a rolling Receives the actual result as inputs and outputs a shape evaluation value. When learning the neural network 1741 For the evaluation function, the actual shape result and the actual rolling result data containing the rolling actual result Si are used as input data, a shape evaluation value to be described later is used as teaching data and a combination thereof is used as the learning data. Thus, the combination of the shape actual result and the rolling actual result and the teaching data as teaching data in the evaluation function teaching data database 1743 saved. The learning execution part can carry out the learning of the neural network in a phase in which learning data has been accumulated to a certain extent.

In addition to the neural network described above 1741 the learning execution part contains a neural network learning control part for the evaluation function 1744 for the evaluation function, an input data creation part 1745 and a teaching data creation part 1746 .

The teaching data creation part 1746 creates teaching data of the shape evaluation value = 1 at the time when the manual operation starts using a manual operation signal of the operator relating to the shape. The teaching data creation section also reports 1746 the time at which the manual operation begins to the input data creation part 1745 . The input data creation part 1745 receives a shape actual result and a rolling actual result at the time the manual operation starts, and uses them as input data. The through the input data creation part 1745 input data created and those created by the teaching data creation part 1746 Teaching data created are stored as a pair of learning data in the evaluation function learning data database 1743 saved.

The teaching data creation part creates accordingly 1746 at the time when the manual operation ends, teaching data of the shape evaluation value = 0 using the operator's manual operation signal relating to the shape. The teaching data creation section also reports 1746 the point in time at which the manual operation ends to the input data creation part 1745 . The input data creation part 1745 receives a shape actual result and a rolling actual result at the time the manual operation ends and uses the same as input data. The through the input data creation part 1745 input data created and those created by the teaching data creation part 1746 Teaching data created are stored as a pair of learning data in the evaluation function learning data database 1743 saved.

If the learning data in the evaluation function learning data database 1743 the neural network learning control part reads to a certain extent (eg 1000 sentences) 1744 for the evaluation function the learning data from the evaluation function learning data database 1743 , it acquires input data and teaching data from the teaching data and applies them to the neural network 1741 for the evaluation function to carry out the learning of the neural network.

When learning the neural network 1741 for the evaluation function in the learning execution part, the neural network 1741 for the evaluation function in the neural network 1740 copied for the evaluation function of the evaluation execution part. Thus the neural network 1740 updated to a new one for the evaluation function. As a result, the tax expenditure is an investigative part 5 and the "control result good / faulty" determination part 6 able to carry out the evaluation using a new neural network for the evaluation function.

In the present embodiment, if the conditions such as the sheet width, the sheet thickness, the steel grade of the material, etc., which represent the controlled objects, the appropriate evaluation functions are considered to be different. Therefore, the learned neural networks, after the learning of all their conditions has been carried out individually, in the evaluation function database DB5 stored as the evaluation functions and used correctly according to the conditions. Furthermore, by considering the sheet width, the sheet thickness, the type of steel, etc. as the rolling actual results, they can also be covered with a single neural network.

The numerical value of the evaluation function obtained from the neural network for the evaluation function may be incorrect during a period until the learning has progressed to a certain extent. Therefore, the evaluation function selection process learning unit 173 the evaluation function taking into account not only the setting evaluation functions A1 to ON , but also select and use the rolling condition.

As described above, the evaluation function setting part 17th the optimal evaluation function for the control output setting part according to the rolling condition 5 and the "control result good / faulty" determination part 6 on.

16 Fig. 12 is a drawing for explaining an overview of the tax expenditure determination part 5 . The tax expenditure determination part 5 contains a rolling phenomenon model 501 and a "shape correction good / faulty" determination part 502 . The tax expenditure determination part 5 receives the actual result data Si from the control object system 1 , the tax operation size S3 from the tax expenditure calculation part 3rd and the information from the output determination database DB3 and applies the control operation size output allowance / inadmissibility data value S4 at the end of the tax operation. With such a configuration, the control output calculates the determination part 5 a change in shape ahead of where in the tax expenditure calculation part 3rd calculated tax operation size S3 to the rolling mill, which is the control object system 1 is output by entering it into the well-known model of the control object facility 1 (Roll phenomenon model 501 in the case of the embodiment in 16 ) is input, and suppresses the control operation quantity output SO when the shape is expected to deteriorate, thereby preventing the shape from deteriorating very much.

The tax operation size becomes more special S3 into the rolling phenomenon model 501 entered one by the control operation size S3 anticipate the change in shape caused, and thereby shape deviation correction quantity forecast data 503 to calculate. On the other hand, the shape deviation correction amount prediction data 503 to the shape detector data Si (shape deviation actual result data 504 at the current time) from the control object system 1 added to shape deviation forecast data 505 to obtain. The shape deviation forecast data 505 are evaluated to enable prediction of how the shape changes when the control operation size S3 to the control object system 1 is issued. Based on the current shape deviation actual result data 504 and shape deviation forecast data 505 determines the "shape correction good / faulty" determination part 502 whether the shape changes in the direction of improvement or in the direction of deterioration to the control operation size output allowance / inadmissibility data value S4 to obtain.

The "shape correction good / faulty" determination part 502 performs a "good / faulty" determination regarding the shape correction specifically in the following manner. First, the “shape correction good / faulty” determination part leads 502 to account for the control priority in the direction of the sheet width using that of the evaluation function setting part 17th set rolling condition corresponding evaluation function a "good / faulty" determination of a change in shape by. For example, the "shape correction good / faulty" leads to the determination part 502 using an evaluation function expressed by the following equation J a "good / faulty" determination regarding the change in shape by In the following equation εfb (i) an actual shape deviation result 504 , is εest (i) a shape deviation prediction 505 , i is a shape detector area, rand is a random number term and is J _Ai one by the evaluation function setting part 17th set evaluation function. $$J=J_{Ai}\left({\epsilon }_{fb}\left(i\right)\right)-J_{Ai}\left({\epsilon }_{est}\left(i\right)\right)+rand$$

When using the evaluation function above J becomes the evaluation function J positive if the shape is good. If the form becomes faulty, the evaluation function J negative. Rand is also the random number term and changes the evaluation result of the evaluation function J on a random number basis. Thus, there is a case in which the evaluation function J positive, even if the shape deteriorates, it becomes possible to learn the relationship between the shape pattern and the control method even in the case where the rolling phenomenon model 501 is not suitable. Here rand, if appropriate, is changed in such a way that the maximum value becomes large when the model of the control object system 1 is uncertain and set to 0 if the control method is learned to some extent and then stable control is to be performed.

The "shape correction good / faulty" determination part 502 calculates the evaluation function J and gives the control operation size output allowance / inadmissibility data value S4 in such a way that when J≥0 the control operation quantity output allowance / inadmissibility data value S4 = 1 (permissible) and when J <0 the control operation quantity output allowance / / inadmissibility data value S4 = 0 (inadmissible) .

• IF (Steueroperationsgrößen-Ausgabezulässigkeits-/-unzulässigkeits-Datenwert S4 = 0) THEN
• Nr. 1-bis-Nr. n-AS-U-Positionsänderungsgrößen-Ausgaben = 0
• Obere erste Zwischen-Verschiebungs-Positionsänderungsgrößen-Ausgabe = 0
• Untere erste Zwischen-Verschiebungs-Positionsänderungsgrößen-Ausgabe = 0
• ELSE
• Nr. 1-bis-Nr. n-AS-U-Positionsänderungsgrößen-Ausgaben = Nr. 1-bis-Nr. n-AS-U-Positionsänderungsgrößen
• Obere erste Zwischen-Verschiebungs-Positionsänderungsgrößen-Ausgabe = obere erste Zwischen-Verschiebungs-Positionsänderungsgröße
• Untere erste Zwischen-Verschiebungs-Positionsänderungsgrößen-Ausgabe = untere erste Zwischen-Verschiebungs-Positionsänderungsgröße und
• ENDIF.

The tax expense suppressing part 4th determines the presence or absence of the output of the control operation quantity output SO to the control object system 1 according to the control operation size output allowance / inadmissibility data value S4 which is the result of an determination by the tax expenditure determination part 5 is. The control operation size output allowance / inadmissibility data value S4 corresponds to No. 1 to No. n-AS-U position change quantity output, an upper first intermediate displacement position change quantity output and a lower first intermediate displacement position change quantity output and is determined by:

• IF (control operation size output allowance / inadmissibility data S4 = 0) THEN
• No. 1 to No. n-AS-U position change quantity outputs = 0
• Upper first intermediate displacement position change quantity output = 0
• Lower first intermediate displacement position change quantity output = 0
• ELSE
• No. 1 to No. n-AS-U position change quantity output = No. 1 to No. n-AS-U position change quantities
• Upper first intermediate displacement position change quantity output = upper first intermediate displacement position change quantity
• Lower first intermediate displacement position change quantity output = lower first intermediate displacement position change quantity and
• ENDIF.

The control execution device 20th performs the above calculation according to the actual result data Si from the control object system 1 (the rolling mill) and outputs the control operation quantity output SO to the control object system 1 (the rolling mill) to thereby perform the shape control.

Now an operation overview of the control method learning device 21 described. The control procedure learning device 21 uses delayed data in the control execution device 20th data used. A time delay Z ^-1 means e ^-TS and indicates that data is at a ^preset time T be delayed. Because the control object system 1 has a time behavior, there is a time delay due to the control operation quantity output SO until the actual result data changes. Therefore, after executing the control operation, learning using the actual result data after the lapse of only the delay time T executed. Since a shape measuring device takes a few seconds to send to the operator after the operation command has been issued AS-U or the first intermediate roller in the shape control to detect a change in shape, the time can be set to T = 2 to 3 seconds or so (since the delay time also changes depending on the type of the shape detector or the rolling speed, the optimal time can be it takes for the change in the control operation end to change shape when set as T).

17th is a drawing for explaining an operation overview of the "control result good / faulty" determination part 6 . Such a “good / faulty” determination evaluation function J _C as expressed by the following equation is used in the “change of shape good / faulty” determination unit 602. $$J_{\mathrm{C.}}=J_{Ai}\left({\epsilon }_{fb}\left(i\right)\right)-J_{Ai}\left({\epsilon }_{last}\left(i\right)\right)$$

Incidentally, in the above equation, εfb (i) represents shape-deviation actual result data contained in the actual result data Si, εlast (i) represents previous shape-deviation actual result data and J _{Ai is} an evaluation function set by the evaluation function setting part. Here is the one through the evaluation function setting part 17th evaluation function J _Ai set beforehand by hand or by the evaluation function learning unit 174 learned evaluation function (learning evaluation function) as the evaluation function J _Ai set. Through the "good / faulty" determination evaluation function J _C it is determined whether a control result is good or incorrect. Further, also in the case where the control operation quantity output allowance / inadmissibility data value S4 , which is the determination result of the tax expenditure determination part 5 is equal to 0 (tax output not allowed), determines that the shape has deteriorated even though the control operation quantity output to the control object plant 1 is 0 in practice.

Here, when the control operation quantity output allowance / inadmissibility data value S4 = 0, it is assumed that the "control result good / faulty" data value S6 = -1. Furthermore, a threshold upper limit LCU and a threshold lower limit LCL set in advance under a threshold condition "LCU ≥ 0 ≥ LCL". If the result of a comparison with the “good / faulty” determination evaluation function J _{C is} “J _C > LCU”, it is assumed that the “control result good / faulty” data value S6 = -1 (the Shape has deteriorated). If LCU ≥ J _C ≥ 0, it is assumed that the “control result good / faulty” data value S6 = 0 (the shape changes in the direction of deterioration). If 0> J _C ≥ LCL, it is assumed that the "control result good / faulty" data value S6 = 1 (the form changes in the direction of improvement). If J _C <LCL, it is assumed that the "control result good / faulty" data value S6 = 0 (the shape has improved).

Here, the “control result good / defective” data value S6 = -1 corresponds to the case in which, since the shape has deteriorated, the output tax output is suppressed. The “control result good / defective” data value S6 = 0 corresponds to the case in which the output tax output is maintained because the shape does not change or the shape has improved. The “control result good / faulty” data value S6 = 1 corresponds to the case in which the output control variable is increased because the shape has changed in the direction of improvement, but may possibly still improve further.

The “good / faulty” determination evaluation function J _C changes when the evaluation function J _Ai deviates. Therefore, it is assumed that the determination result of the "control result good / faulty" data value S6 deviates. Therefore, the control method learning device guides 21 a determination regarding the "control result good / faulty" data value S6 with regard to each evaluation function set in advance.

Now there is an overview of the learning data creation part 7 described. As in 3rd shown, the learning data creation part creates 7 on the basis of the determination result (the "control result good / faulty" data value S6 ) from the "Control result good / faulty" determination part 6 Teaching dates S7a regarding that in the control rule learning part 11 used neural network 111 from the control operation end operation command S2 , the control operation size S3 and the determination result of the control output suppressing part (control operation quantity output allowance / inadmissibility data value S4 ).

The teaching dates S7a in this case the AS-U actuation level 301 and the first intermediate degree of operation 302 which is the output from the output layer of the neural network 111 correspond and which both in 7 are shown. The learning data creation part 7 created using the control operation end operation command S2 (AS-U degree of actuation 301 , first intermediate degree of actuation 302 ), which is the output of the neural network 101 and the No. 1 to No. n-AS-U position change quantity output, the upper first intermediate displacement position change quantity output and the lower first intermediate displacement position change quantity output, which is the control operation quantity output SO is, teaching data S7a regarding that in the control rule learning part 11 used neural network 111 .

When describing the operation overview of the learning data creation part 7 is the relationship between the data and sign of the respective parts in the tax output calculation part 3rd in 10th in 18th shown. Here is the control operation end operation command S2 which is the output of the neural network 101 is, typically through the AS-U Degree of activity 301 shown. The control operation end operation command S2 will be described on the assumption that data on the positive side of the degree of operation is OPref, data on the negative side of the degree of operation is OMref, the degree of operation generated on a random number basis from the control operation failure generation part 16 is an operation degree random number Oref, a conversion gain G and the control operation size output SO Cref is. So here are for simplification than the output from the output layer of the neural network 101 in the control rule execution part 10th , the positive side of the operation degree and the negative side of the operation degree, and the operation degree generated on the random number basis from the control operation failure generation part 16 defined as random number of actuation levels. Furthermore, the control operation quantity output SO defined as an operation command value with respect to the control operation end.

19th shows a processing stage and processing content in the learning data creating part 7 . Here, if according to the sign's promise in 18th are described, the operation command value Cref from an equation (6) in the first processing stage 71 determined. $$\text{C.}ref=G\cdot \left(\text{OP}ref-\text{OM}ref+\text{OR}ref\right)$$

$$\text{C}\text{'}ref=\text{C}ref$$

$$\begin{array}{c}\text{IF C}ref>0\text{THEN C}\text{'}ref=\text{C}ref+\mathrm{\Delta }\text{C}ref\\ \text{IF C}ref>0\text{THEN C}\text{'}ref=\text{C}ref-\mathrm{\Delta }\text{C}ref\end{array}\}$$

In the next processing stage 72 becomes the operation command value Cref according to the "control result good / faulty" data value S6 corrected to C'ref to obtain. Specifically, the operation command value Cref by an equation (7) when the “control result good / faulty” data value S6 = -1, an equation (8) when the “control result good / faulty” data value S6 = 0, and an equation (9) , if the "control result good / faulty" data value S6 = 1, set to have a correction value C'ref the same assumes. $$\begin{array}{c}\text{IF C}ref>0\text{THEN C}\text{'}ref=\text{C.}ref-\mathrm{\Delta }\text{C.}ref\\ \text{IF C}ref>0\text{THEN C}\text{'}ref=\text{C.}ref+\mathrm{\Delta }\text{C.}ref\end{array}\}$$

$$\text{C.}\text{'}ref=\text{C.}ref$$

$$\begin{array}{c}\text{IF C}ref>0\text{THEN C}\text{'}ref=\text{C.}ref+\mathrm{\Delta }\text{C.}ref\\ \text{IF C}ref>0\text{THEN C}\text{'}ref=\text{C.}ref-\mathrm{\Delta }\text{C.}ref\end{array}\}$$

$$\mathrm{\Delta }\text{O}ref=\frac{1}{2}\left(\frac{\text{C}\text{'}ref}{G}-\left(\text{OP}ref-\text{OM}ref\right)\right)$$

In the processing stage 73 becomes the operation degree correction quantity ΔCref by equations (10) and (11) according to the corrected operation command value C'ref determined. $$\text{C.}\text{'}ref=G\cdot \left(\left(\text{OP}ref+\mathrm{\Delta {\rm O}}ref\right)-\left(\text{OM}ref-\mathrm{\Delta }\text{O}ref\right)\right)$$

$$\mathrm{\Delta }\text{O}ref=\frac{1}{\mathrm{2nd}}\left(\frac{\text{C.}\text{'}ref}{G}-\left(\text{OP}ref-\text{OM}ref\right)\right)$$

In the processing stage 74 become the teaching data OP'ref and OM'ref for the neural network 111 determined by an equation (12). $$\begin{array}{c}\text{OP}\text{'}ref=\text{OP}ref+\mathrm{\Delta }\text{O}ref\\ \text{OM}\text{'}ref=\text{OM}ref-\mathrm{\Delta }\text{O}ref\end{array}\}$$

Thus calculated as in 18th shown, the learning data creation part 7 actually the operation command value correction value C'ref according to the "control result good / faulty" data value S6 which the determination result in the "control result good / faulty" determination part 6 is from which to the control object system 1 issued operation command value Cref . Specifically, in the case where the "control result good / faulty" data S6 = 1, this corresponds to the case where the control direction is OK, but it is determined that the control output is insufficient, and the operation command value becomes ΔCref increased in the same direction. In the case in which the "control result good / faulty" data value S6 = -1 is reversed, this corresponds to the case in which the control direction is determined to be incorrect and the control command value is changed ΔCref decreased in an opposite direction. Because the conversion gain G The correction variable can be set in advance and is therefore already known ΔOref be determined if the values on the positive and negative side of the degree of actuation are known. Here will ΔCref by determining an appropriate value in advance by simulation or the like. According to the above procedures, those in the control rule learning part 11 used teaching data OP'ref and OM'ref can be determined by the above equation (12).

Incidentally, although the above description is based on the simple case studies in 19th the AS-U degree of actuation 301 regarding the AS-U No. 1 to No. n and the first intermediate degree of operation 302 regarding the displacement of the upper first intermediate roller and the displacement of the lower first intermediate roller are all practically carried out and are as the teaching data (AS-U operation degree teaching data and first-intermediate operation degree teaching data) of that in the control rule learning part 11 used neural network 111 Are defined.

20th shows one in the learning data database DB2 saved data example. Combinations of a large number of input data S8a and teaching dates S7a are required to the neural network 111 to learn. Thus, in the learning data creation part 7 created teaching data S7a (AS-U degree of activity teaching data, first intermediate degree of activity) with the time-delayed data S8a by the control execution device 20th in the control rule execution part 10th input data entered S1 (standardized form deviation 201 and shape deviation level) and they are as a pair of learning data S11 , which in turn is in the learning data database DB2 are saved.

Incidentally, while the different databases DB1 , DB2 , DB3 , DB4 , and DB5 in the plant control device in 3rd were used in 20th a configuration of a neural network management table TB to manage and operate the respective databases DB1 , DB2 , DB3 and DB4 shown on a link basis. The administration table TB contains a specification management table. The administration table is special TB according to ( B1 ) the sheet width, ( B2 ) of the steel grade and the evaluation functions A1 to ON divided over the control priority with regard to the specification. As ( B1 ) The sheet width, for example, four divisions of 3 feet wide, 1 meter wide, 4 feet wide and 5 feet wide are used, and as the steel grade 10th Subdivisions of steel grades ( 1 ) to ( 10th ) or so used. Further, regarding the evaluation functions of the controller, N (N is the set number of evaluation functions, and in the present embodiment, N = 6) types are used. In this case the number of divisions 80 and 240 neural networks are properly used and used according to the rolling conditions.

The neural network learning control part 112 links the learning data, which is the combination of the input data and the teaching data as in 20th are shown, according to the neural network management table TB in 21 with the corresponding neural network no. and is in such a learning data database DB2 as in 22 shown saved.

Every time the control execution device 20th the form control on the control object system 1 carried out, the learning data are created in two sentences according to the evaluation function. The reason for this is that N types of teaching data are created because the “control result good / faulty” determination is carried out using the N evaluation functions via the priority of the control for the same input data and the same control output. When the teaching data is accumulated to a certain extent (e.g. 200 sentences) or in the learning data database DB2 are accumulated, the neural network learning control part causes 112 learning the neural network 111 .

A variety of neural networks have been created according to such a management table TB as in 21 shown in the control rule database DB1 saved. The neural network learning control part 112 names the number of the neural network that requires learning and the neural network selection unit 113 takes the corresponding neural network from the control rule database DB1 and put the same on the neural network 111 . The neural network learning control part 112 instructs the input data creation unit 114 and the teaching data creation unit 115 the input data and the teaching data, which correspond to the corresponding neural network, from the learning data database DB2 and uses the same to learn the neural network 111 out. Incidentally, various methods have been proposed as the learning method of the neural network, and any of them can be used.

After completing the learning of the neural network 111 writes the neural network learning control part 112 the neural network 111 , which is the learning result, back to the position of the corresponding neural network no. the control rule database DB1 where the learning is complete.

Learning can be done in all 21 defined neural networks at a point in time at prescribed time intervals (for example daily), or at this point in time only neural networks in the neural network number can be used, in which new learning data has been accumulated to a certain extent (for example 100 sets) be learned.

1. 1.) Ein Referenz-Formmuster und eine diesem entsprechende Steueroperation werden im Voraus einzeln eingestellt, die Kombination des Formmusters und der Steueroperation wird gelernt, ohne ein Steueroperations-Verfahren zu lernen, und die Steueroperation wird unter Verwendung derselben durchgeführt,
2. 2.) da eine neue Steuerungsregel nicht vorab vorausberechnet werden kann und eine vollkommen unberechenbare Steuerungsregel optimal gemacht werden kann, wird das Steueroperationsende aufs Geratewohl ausgeführt, um bei Betrachtung desselben ein darauf bezogenes Steuerungsergebnis zu ermitteln, und
3. 3.) eine Auswertungsfunktion zum Ermitteln der Steuerungspriorität bezüglich eines gesteuerten Objekts wird so eingestellt, dass sie zu der Ansicht des Bedieners passt und zu einem Handbetätigungsverfahren des Bedieners gemäß dem Zustand des gesteuerten Objekts passt.

Thus, the following can be realized without the shape of the rolling mill, which is the control object system 1 is very disturbing:

1. 1.) A reference shape pattern and a control operation corresponding thereto are individually set in advance, the combination of the shape pattern and the control operation is learned without learning a control operation method, and the control operation is performed using the same,
2. 2.) since a new control rule cannot be calculated in advance and a completely unpredictable control rule can be made optimal, the control operation end is carried out at random to determine a control result related to it, and
3. 3.) an evaluation function for determining the control priority with respect to a controlled object is set in such a way that it fits the view of the operator and fits a manual operation method of the operator according to the state of the controlled object.

Incidentally, this is in the control execution device 20th used neural network in the control rule database DB1 saved. However, if the stored neural network is only subjected to the initial processing with a random number, the learning of the neural network proceeds and takes time until suitable control becomes possible. Therefore, when developing the control part for the control object system 1 learning the control rule in advance by simulation based on the control model of the control object system 1 , which is known at this time, and the neural network in which the learning is completed in a simulator is stored in the database, whereby it is possible to control the performance to some extent from the start of the control object system.

Furthermore, since the learning of the neural network is carried out using the evaluation function matching the operator's manual operation method, it is not necessary that the operator perform the manual operation on the change in the controlled object due to the control output. It is possible to reduce loads imposed on the operator and to improve control accuracy and operational efficiency.

The embodiment described above contains points shown below. However, the items included in the embodiment are not limited to the items shown below.

The control device of the present disclosure is a control device that controls a controlled object. The control device includes a control execution device that applies a control output to the controlled object according to a control rule given to it, a control method learning device that evaluates the control output applied to the controlled object using a specified evaluation function by using an evaluation result of the same learning data, learns the learning data to thereby create the control rule and applies the control rule to the control execution device, and an evaluation function setting part that includes a plurality of evaluation functions in advance selects one of the many evaluation functions based on a control state for the controlled object and notifies the selected evaluation function of the control method learning device.

According to this configuration, since the control output is applied to the controlled object according to the control rule created by learning the learning data after using the evaluation result of the evaluation for the control output by the evaluation function selected based on the control state, the control can be expected based on a suitable "good / faulty" determination can be carried out with regard to the control result.

In addition, according to the present disclosure, the evaluation function setting part calculates and selects an “evaluation function good / faulty” determination index for each of the many evaluation functions based on the control state for the controlled object and manual operation by an operator, and selects it based on the “evaluation function good / defective ”determination index an evaluation function specified for the control method learning device. According to this configuration, by using the relationship between the manual operation by the operator and the control state for the controlled object, it is likely that the evaluation function in which the control intended by the operator is highly rated is selected.

Further, according to the present disclosure, the evaluation function setting part calculates an evaluation value of the evaluation function both at the time when the operator starts the manual operation and at the time at which the operator ends the manual operation, and he uses the evaluation value to calculate the “evaluation function good / faulty” determination index. According to this configuration, since the operator starts manual operation when it is determined that the shape of a sheet has deteriorated during a rolling operation, and manual operation continues until it is determined that the shape has improved, the intention of the Operator obtained from the evaluation value at that time.

Further, according to the present disclosure, the evaluation function setting part calculates an evaluation value a of the evaluation function at the time when the operator starts the manual operation and an evaluation value b of the evaluation function at the time when the operator ends the manual operation, and calculates the “ Evaluation function good / faulty ”-determination index as (ab) / b. According to this configuration, even if a calculation method is different for several evaluation functions, it is possible to compare “evaluation function good / faulty” indices with one another.

Furthermore, according to the present disclosure, the evaluation function serves to output the evaluation result with the control output to the controlled object and actual result data of the controlled object, in which a control result of the control output is reflected, as inputs. The evaluation function setting part learns learning data based on the manual operation by the operator, the control output to the controlled object, and the actual result data of the controlled object, to thereby create the evaluation function. According to this configuration, since the manual operation of the operator is used, it is possible to create an evaluation function in which the operator's intention is reflected.

Furthermore, according to the present disclosure, the evaluation function setting part learns data based on the control output to the controlled object and the actual result data of the controlled object at the time when the operator starts the manual operation and at the time when the operator starts the operation Manual operation ended in order to create the evaluation function. According to this configuration, since the operator starts the manual operation when it is determined that the shape of the sheet has deteriorated during the rolling operation and continues the manual operation until it is determined that the shape has improved, an evaluation function is possible Create, which carries out an evaluation close to the evaluation of the operator with an evaluation value reflecting the evaluation of the operator as learning data.

Furthermore, according to the present disclosure, the evaluation function setting part generates learning data with an evaluation value at the time at which the operator starts the manual operation as a predefined value c, it generates learning data with an evaluation value at the time at which the operator ends the manual operation , as a predefined value d and it learns the learning data in order to thereby create the evaluation function. According to this configuration, since the operator starts manual operation when it is determined that the shape of the sheet has deteriorated during the rolling operation, and manual operation continues until it is determined that the shape has improved, the intention of the Operator obtained from the evaluation value at that time.

Furthermore, according to the present disclosure, the control execution device includes a control rule execution part, which applies a control output to the controlled object according to a combination of the actual result data of the controlled object and a control operation, a control output determination part, which determines using the evaluation function, whether the control output output from the control rule execution part is applicable, and, if it is determined that the control output is not to be applied, notifies the control method learning device of the unsuitability of the combination of the actual result data and the control operation, and a control output suppressing part which, when the tax expenditure determination part has determined that the tax expenditure is not applicable, prevents the tax expenditure from being output to the controlled object. The control method learning device includes a “control result good / faulty” determination part, which, when the control execution device actually outputs the control output to the controlled object, after a time delay until the control output is reflected in the actual result data of the controlled object Using the evaluation function set by the evaluation function setting part, a determination is carried out regarding whether the control result is free of errors / inaccuracy as to whether the actual result data improve or deteriorate due to the control output, a learning data creation part which, using the control result good / faulty ”determination part determines freedom from errors / faultiness of the control result and the tax output for teaching data, and a control rule learning part which learns the actual result data and the teaching data as learning data. With the learning by the control method learning device, a combination of individual actual Result data and control operations for a large number of controlled objects according to the state of the control object system. The combination of the actual result data and the control operations thus obtained is used as a defined combination of the actual result data and a control operation of the control object system in the control rule execution part.

Furthermore, the plant control device of the present disclosure may actually be implemented as a computer system. In this case, however, a plurality of program groups are formed in the computer system.

These program groups are, for example, programs for realizing the processing of a control execution device: a control rule execution program which uses a control output according to a defined combination of actual result data of a control object system and a control operation, a control output determination program which determines whether the control output outputted by the control rule execution program is permitted, and the control method learning device reports that the actual result data and the control operation are incorrect, and a control output suppression program that, when the control output determination program has determined that the actual result data of the Control object system deteriorate when the control output is output to the control object system, prevents the control output from being output to the control object system, and
Programs for realizing the processing of the control method learning device: a “control result good / faulty” determination program for, when the control execution device actually outputs the control output to the control object system, realizing the processing of a “control result good / faulty” determination to after a time delay until a control effect is shown in the actual result data, a determination as to whether the control result is error-free / faulty about whether the actual result data improve or deteriorate compared to that before the corresponding control, a learning data creation program which using the correctness / erroneousness of the control result in the "control result good / defective" determination program and the control output wins teaching data, and a control rule learning program which obtains the actual result data and the teaching data learns as learning data. Further, with the learning by the control method learning device, a combination of individual actual result data and a control operation for a plurality of controlled objects is obtained in accordance with the state of the control object system, and the combination of the actual result data and the control operation thus obtained becomes a defined combination of actual result data and a control operation of the control object system are used in the control rule execution program.

Incidentally, it is necessary to set an initial value of a neural network when the device of the present invention is applied to a concrete plant. Regarding this point, however, before executing control on the control object system by simulation using a control model of a control object system, it is desirable to create a combination of actual result data and a control operation and a period of time for learning the combination of the actual result data and to shorten the control operation in the control object system.

Industrial applications

The present invention relates to a control method and a section of a rolling mill, which is one of the rolling mills, for example, and there is no particular problem in practical use.

QUOTES INCLUDE IN THE DESCRIPTION

This list of documents listed by the applicant has been generated automatically and is only included for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent literature cited

• JP 2804161 [0008]
• JP 4003733 [0008]

Claims (9)

A control device which controls a controlled object, comprising: a control execution device that applies a control output to the controlled object according to a control rule given to it; a control method learning device that evaluates the control output applied to the controlled object using a specified evaluation function by using an evaluation result of the same learning data, learns the learning data to thereby create the control rule, and applies the control rule to the control execution device; and an evaluation function setting part that includes a plurality of evaluation functions in advance, selects one of the many evaluation functions based on a control state for the controlled object, and notifies the selected evaluation function to the control method learning device. Control device after Claim 1 , wherein the evaluation function setting part calculates an "evaluation function good / faulty" determination index for each of the many evaluation functions based on the control state for the controlled object and manual operation by an operator, and one based on the "evaluation function good / faulty" determination index selects the evaluation function specified for the control method learning device. Control device after Claim 2 , wherein the evaluation function setting part calculates an evaluation value of the evaluation function both at the time at which the operator starts the manual operation and at the time at which the operator ends the manual operation, and using the evaluation value the “evaluation function good / faulty” -Determination index calculated. Control device after Claim 3 , wherein the evaluation function setting part calculates an evaluation value a of the evaluation function at the time at which the operator starts the manual operation and an evaluation value b of the evaluation function at the time at which the operator ends the manual operation, and the "evaluation function good / faulty" -Determination index calculated as (ab) / b. Control device after Claim 1 , wherein the evaluation function serves to output the evaluation result with the control output to the controlled object and actual result data of the controlled object, in which a control result of the control output is reflected, as inputs, and wherein the evaluation function setting part is based on the manual operation by the Operator who learns control output to the controlled object and the actual result data of the controlled object learning data in order to thereby create the evaluation function. Control device after Claim 5 wherein the evaluation function setting part learns learning data based on the control output to the controlled object and the actual result data of the controlled object at the time when the operator starts the manual operation and the time when the operator ends the manual operation. to thereby create the evaluation function. Control device after Claim 6 , wherein the evaluation function setting part generates learning data with an evaluation value at the time when the operator starts the manual operation as a predefined value c, learning data with an evaluation value at the time at which the operator ends the manual operation as a predefined value d generated and learns the learning data in order to create the evaluation function. Control device after Claim 1 , wherein the control execution device comprises a control rule execution part which applies a control output to the controlled object according to a combination of the actual result data of the controlled object and a control operation, a control output determination part which determines using the evaluation function whether the result of the control rule -Execution part outputting control output is applicable, and when it is determined that the tax output is not to be applied, the control method learning device notifies the unsuitability of the combination of the actual result data and the control operation, and a control output suppressing part, which if the control output determination part has determined that the control output is not to be applied, prevents the control output from being output to the controlled object, and wherein the control method learning device includes a "control result good / faulty" determination part, which if the S expensive execution device actually the control output to the controlled object outputs, using the evaluation function set by the evaluation function setting part after a time delay until the control output is reflected in the actual result data of the controlled object, a determination as to whether the control result is free of errors / defective in whether the actual result data differs from the control output improve or worsen, carries out a learning data creation part, which obtains teaching data using the error-free / defective determination of the control result and the control output, determined by the “control result good / faulty” determination part, and a control rule learning part, which the actual result data and the Teaching data as learning data is learned, includes, and learning by the control method learning device is a combination of individual actual result data and a control operation for a plurality of controlled objects according to the state of the control object ject system and the combination of actual result data and the control operation thus obtained is used as a defined combination of actual result data and a control operation of the control object system in the control rule execution part. A control method for controlling a controlled object which is executed by a computer, comprising: applying control output to the controlled object according to a given control rule; evaluating the control output applied to the controlled object using a specified evaluation function; creating learning data by using an evaluation result thereof; learning the learning data to create the control rule; and selecting and specifying one of a plurality of evaluation functions included in advance based on a control state for the controlled object.

Images